When an AI agent's permission should expire
What should invalidate an earlier approval when the evidence changes?
A support agent needs a person's approval to refund $80. The approval is granted. Then the agent waits in a queue. While it waits, the order changes, or the customer record changes, or someone else already issued a refund. The approval still exists. It no longer fits the situation.
Checking the agent's identity or role does not help here. The system has to check whether this specific refund is still covered.
Make permission specific enough to check
Bind the approval to the things that would make it wrong if they changed: the order, the customer, the amount, the destination, the version of the evidence it was granted on, and an expiry time. A material change to any of them should require a new decision.
A check inside the agent's planning loop still leaves a gap between checking and acting. Where the executing service can enforce the conditions atomically — a version check at the moment of execution, for instance — it should.
Revocation needs an execution boundary
Temporary credentials limit how long access lasts. They do not prove an approved action is still appropriate, and revoking access cannot undo a refund that has already gone out. The audit trail has to distinguish a refund requested from a refund confirmed, including the uncertain state after a timeout.
Try the example
A small Python example with a fake payment adapter and six local tests shows the pattern. It makes no real payments and no network calls, and it does not validate the research prototype.